The agent standards stack went from 3 projects to 10 in a year: Anthropic, OpenAI, Hugging Face and Alibaba hand four core projects to neutral governance
The real story is not four more open-source repositories. It is four directly competing companies, Anthropic, OpenAI, Hugging Face and Alibaba, carving out a piece each and handing it to a foundation nobody controls. After a year of competition the industry ran into an awkward fact: if every vendor's skill format, model interface and sandbox protocol differs, customers cannot compose agents and the market stays small.
Pin the source first. The primary material is a Linux Foundation press release published October 22, 2026 titled Agentic AI Foundation Announces Four New Projects: Agent Skills, Open Responses, OpenEnv, and OpenSandbox, released at AGNTCon + MCPCon North America, held October 22 to 23 at the San Jose McEnery Convention Center. Every figure, project definition and quote below comes from that release.
Project one: Agent Skills, contributed by Anthropic. The release defines it as an open format for giving agents new capabilities and domain expertise, packaging instructions, scripts and resources in a reusable form that agents load on demand, which makes specialised know-how portable across agent products and platforms. David Soria Parra, co-creator and lead maintainer of MCP at Anthropic, said they launched Agent Skills a year ago to give developers and enterprises an open standard, because without the right context even the most intelligent agents fall short.
Project two: Open Responses, contributed by OpenAI. It is an open specification for building multi-provider, interoperable LLM interfaces based on the OpenAI Responses API, defining a shared schema, consistent streaming events and extensible tooling so developers describe requests and outputs once and run them across model providers with minimal translation. Steve Coffey of OpenAI said the project joined AAIF to keep pursuing its mission of helping developers build great products through simple, clean, unified interfaces.
Project three: OpenEnv, contributed by Hugging Face. It is an open framework and community hub for creating RL environments that are interoperable, plug-and-play, easy to build, easy to share and scalable, delivered as a containerised service exposing a familiar agent interface, MCP, and Gym, the training protocol training libraries use. Ben Burtenshaw of Hugging Face said RL environments are a fundamental part of an open stack because they teach agents the capabilities they will use in the real world, naming Microsoft, Reflection, Meta-PyTorch, Fleet, Mercor, RadixArk, Modal and Prime Intellect among participants.
Project four: OpenSandbox, contributed by Alibaba. It is a general-purpose sandbox platform for AI applications offering multi-language SDKs, unified sandbox APIs and both Docker and Kubernetes runtimes, covering coding agents, GUI agents, agent evaluation, AI code execution and reinforcement learning training. It includes a sandbox protocol defining lifecycle and execution APIs so teams can extend it with custom runtimes. Xiaobin Xu, head of development infrastructure at Alibaba, put it plainly: while OpenSandbox was created at Alibaba, we do not want its future to be defined by a single company.
The four projects are four layers, and that layering carries the information. The release says they address four layers of the agentic stack: how agents gain new skills with Agent Skills, how applications talk to models across providers with Open Responses, how agents are trained and evaluated with OpenEnv, and where agents run safely with OpenSandbox. Stack them and you get close to a complete agent runtime, and the contributors are four companies that compete with each other.
Get the numbers right. AAIF went from three launch projects to 10 in its first year. The original three were MCP, goose and AGENTS.md in December 2025, and the four new projects now sit alongside the Agent2Agent protocol, Agent Router and agentgateway to make 10, all under AAIF's open, vendor-neutral governance. The release also notes AAIF surpassed 300 members, which is to say governance expansion and project expansion happened in step.
How the foundation frames it. Manik Surtani, CTO and co-founder of AAIF, said: A year ago AAIF launched with three projects and a bet that agentic AI needed a neutral home to thrive. Today we host 10. Agent Skills, Open Responses, OpenEnv and OpenSandbox each cover a different part of that stack, from how agents gain capabilities to where they run, and four different companies chose to contribute them to neutral governance. That is the job of this foundation: putting the community first by giving it one place to build shared standards, protocols and infrastructure, with no single company controlling the outcome.
Why this matters more to developers than to vendors. Handing over an interface looks like giving up control in the short run, but developers get a predictable list: skill format, model interface, training environment and execution sandbox each have an owner, a place to file issues, and a roadmap not dictated by one company's product cycle. For an enterprise mid-selection that is steadier than any single vendor promise, because promises move with strategy while governance does not turn on a dime.
Where the boundaries are. First, this is a Linux Foundation release plus executive statements, so it is first-party, but maturity varies wildly: MCP is a de facto standard while some of the four are newly opened and still early. Second, the release gives no adoption numbers, performance data or security audit results, so it cannot support a claim that any project is better. Third, neutral governance is not the same as implementation quality; code quality, maintenance cadence and compatibility still have to be checked in each repository. Fourth, this article records only what was officially published through October 26, 2026.
🤔 Frequently Asked Questions
What is AAIF and how does it relate to the Linux Foundation?
The Agentic AI Foundation sits under the Linux Foundation as the neutral home for open standards, protocols and infrastructure in agentic AI. Its founding projects were MCP, goose and AGENTS.md, launched in December 2025. It governs the core standards, protocols and frameworks that let agents interoperate across platforms, operating under open, vendor-neutral principles.
What does each of the four projects solve?
By the foundation's own layering: Agent Skills handles how agents gain new skills and domain expertise by packaging instructions, scripts and resources into a reusable, on-demand format; Open Responses handles how applications talk across model providers with a shared schema, consistent streaming events and extensible tooling; OpenEnv handles how agents are trained and evaluated through an interoperable, plug-and-play RL environment framework; OpenSandbox handles where agents run safely, with multi-language SDKs, unified sandbox APIs and Docker or Kubernetes runtimes.
Why would four competitors give these projects away?
The release frames neutral governance as letting the community shape what comes next. The stated motives are readable in the quotes: Anthropic wanted anyone building with skills to shape what follows, and Alibaba was bluntest with we do not want its future to be defined by a single company. The common thread is that if every vendor defines its own format, customers cannot compose anything and nobody's market grows.
What does this mean in practice for teams building agent products?
The practical move is to audit your stack against those four layers and ask, per layer, whether you built it yourself or can align to an open standard. Align skill formats with Agent Skills, consider Open Responses for model interfaces, look at OpenEnv for training and evaluation, and evaluate OpenSandbox or peers for execution. The payoff is controllable migration cost when you change models, clouds or sandboxes. Treat it as a direction, not a verdict: check each project's maturity in its own repository.
🛠️ Recommended Tools
- AI Agent Interoperability because all four projects exist to let agents from different vendors compose, so map your interoperability boundary first.
- AI Agent Sandbox since OpenSandbox lives on this layer, so clarify what a sandbox must provide before choosing one.
- AI Prompt Templates since Agent Skills is essentially reusable packaging of instructions and resources, and templating is the easiest first step.
- Regex Tester and regex remains the handiest debug tool when parsing structured model output against a schema.
Summary
The Linux Foundation's Agentic AI Foundation announced four new projects on October 22, 2026 at AGNTCon + MCPCon North America: Agent Skills from Anthropic, an open format packaging instructions, scripts and resources into reusable, on-demand units so expertise is portable across agent products; Open Responses from OpenAI, an open specification for multi-provider LLM interoperability based on the Responses API, defining a shared schema, consistent streaming events and extensible tooling; OpenEnv from Hugging Face, an open RL environment framework and community hub exposing an agent interface, MCP and Gym through a containerised service; and OpenSandbox from Alibaba, a general-purpose sandbox platform with multi-language SDKs, unified sandbox APIs and Docker or Kubernetes runtimes covering coding agents, GUI agents, evaluation, code execution and RL training. The four map onto four layers of the agentic stack. AAIF moved from three projects to 10 in its first year: the original three were MCP, goose and AGENTS.md in December 2025, and they now sit with A2A, Agent Router, agentgateway and the four additions, all under open, vendor-neutral governance, with membership passing 300. The release provides no adoption figures, performance metrics or security audit results.
Sources: Linux 基金会官方新闻稿：Agentic AI Foundation Announces Four New Projects（2026-10-22）
Linux 基金会官方新闻稿：Agentic AI Foundation Surpasses 300 Members（2026-10）